Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
A Logical Characterization of Iterated Admissibility Joseph Y. Halpern and Rafael Pass Computer Science Department, Cornell University, Ithaca, NY, 14853, U.S.A. e-mail: [email protected], [email protected] June 23, 2009 Abstract Brandenburger, Friedenberg, and Keisler provide an epistemic characterization of iterated admissibility (i.e., iterated deletion of weakly dominated strategies) where uncertainty is represented using LPSs (lexicographic probability sequences). Their characterization holds in a rich structure called a complete structure, where all types are possible. Here, a logical charaacterization of iterated admisibility is given that involves only standard probability and holds in all structures, not just complete structures. A stronger notion of strong admissibility is then defined. Roughly speaking, strong admissibility is meant to capture the intuition that “all the agent knows” is that the other agents satisfy the appropriate rationality assumptions. Strong admissibility makes it possible to relate admissibility, canonical structures (as typically considered in completeness proofs in modal logic), complete structures, and the notion of “all I know”. 1 Introduction Admissibility is an old criterion in decision making. A strategy for player i is admissible if it is a best response to some belief of player i that puts positive probability on all the strategy profiles for the other players. Part of the interest in admissibility comes from the observation (due to Pearce [1984]) that a strategy σ for player i is admissible iff it is not weakly dominated; that is, there is no strategy σ 0 for player i that gives i at least as high a payoff as σ no matter what strategy the other players are using, and sometimes gives i a higher payoff. It seems natural to ignore strategies that are not admissible. But there is a conceptual problem when it comes to dealing with iterated admissibility (i.e., iterated deletion of weaklhy dominated strategies). As Mas-Colell, Whinston, and Green [1995, p. 240] put in their textbook when discussing iterated deletion of weakly dominated strategies: [T]he argument for deletion of a weakly dominated strategy for player i is that he contemplates the possibility that every strategy combination of his rivals occurs with positive probability. However, this hypothesis clashes with the logic of iterated deletion, which assumes, precisely, that eliminated strategies are not expected to occur. Brandenburger, Friedenberg, and Keisler [2008] (BFK from now on) resolve this paradox in the context of iterated deletion of weakly dominated strategies by assuming that strategies are not really eliminated. Rather, they assumed that strategies that are weakly dominated occur with infinitesimal (but nonzero) probability. (Formally, this is captured by using an LPS—lexicographically ordered probability sequence.) They define a notion of belief (which they call assumption) appropriate for their setting, and show that strategies that survive k rounds of iterated deletion are ones that are played in states where there there is kth-order mutual belief in rationality; that is, everyone assume that everyone assumes . . . (k − 1 times) that everyone is rational. However, they prove only that their characterization of iterated admissibility holds in particularly rich structures called complete structures (defined formally in Section 4), where all types are possible. Here, we provide an alternate logical characterization of iterated admissibility. The characterization simply formalizes the intuition that an agent must consider possible all strategies consistent with the rationality assumptions he is making. Repeated iterations correspond to stronger rationality asumptions. The characterization has the advantage that it holds in all structures, not just complete structures, and assumes that agents represent their uncertainty using standard probability meaures, rather than LPS’s or nonstandard probability measures (as is done in a characterization of Rajan [1998]). Moreover, while complete structures must be uncountable, we show that our characterization is always satisfible in a structure with finitely many states. In an effort to understand better the role of complete structures, we consider strong admissibility. Roughly speaking, strong admissibility is meant to capture the intuition that “all the agent knows” is that the other agents satisfy the appropriate rationality assumptions. We are using the phrase “all agent i knows” here in the same sense that it is used by Levesque [1990] and Halpern and Lakemeyer [2001]. We formalize strong admissibility by requiring that the agent ascribe positive probability to all formulas consistent with his rationality assumptions. (This admittedly fuzzy description is made precise in Section 3.) We give a logical characterization of iterated strong admissibility and show that a strategy σ survives iterated deletion of weakly dominated strategies iff there is a structure and a state where σ is played and the formula characterizing iterated strong admissibility holds. While we can take 1 the structure where the formula holds to be countable, perhaps the most natural structure to consider is the canonical structure, which has a state corresponding to very satisfiable collection of formulas. The canonical structure is uncountable. We can show that the canonical structure is complete in the sense of BFK. Moreover, under a technical assumption, every complete structure is essentially canonical (i.e., it has a state corresponding to every satisfiable collection of formulas). This sequence of results allows us to connect (iterated admissibility), complete structures, canonical structures, and the notion of “all I know”. 2 Characterizing Iterated Deletion We consider normal-form games with n players. Given a (normal-form) n-player game Γ, let Σi (Γ) denote the strategies of player i in Γ. We omit the parenthetical Γ when it is clear from context or ~ = Σ1 × · · · × Σn . irrelevant. Let Σ Let L1 be the language where we start with true and the special primitive proposition RAT i and close off under modal operators Bi and hBi i, for i = 1, . . . , n, conjunction, and negation. We think of Bi ϕ as saying that ϕ holds with probability 1, and hBi iϕ as saying that ϕ holds with positive probability. As we shall see, hBi i is definable as ¬Bi ¬ if we make the appropriate measurability assumptions. To reason about the game Γ, we consider a class of probability structures corresponding to Γ. A probability structure M appropriate for Γ is a tuple (Ω, s, F, PR1 , . . . , PRn ), where Ω is a set of states; s associates with each state ω ∈ Ω a pure strategy profile s(ω) in the game Γ; F is a σ-algebra over Ω; and, for each player i, PRi associates with each state ω a probability distribution PRi (ω) on (Ω, F) such that, (1) for each strategy σi for player i, [[σi ]]M = {ω : si (ω) = σi } ∈ F, where si (ω) denotes player i’s strategy in the strategy profile s(ω); (2) PRi (ω)([[si (ω)]]M ) = 1; (3) for each probability measure π on (Ω, F), and player i, [[π, i]]M = {ω : Πi (ω) = π} ∈ F; and (4) PRi (ω)([[PRi (ω), i]]M ) = 1. These assumptions essentially say that player i knows his strategy and knows his beliefs. The semantics is given as follows: • (M, ω) |= true (so true is vacuously true). • (M, ω) |= RAT i if si (ω) is a best response, given player i’s beliefs on the strategies of other players induced by PRi (ω). (Because we restrict to appropriate structures, a players expected utility at a state ω is well defined, so we can talk about best responses.) • (M, ω) |= ¬ϕ if (M, ω) 6|= ϕ. • (M, ω) |= ϕ ∧ ϕ0 iff (M, ω) |= ϕ and (M, ω) |= ϕ0 • (M, ω) |= Bi ϕ if there exists a set F ∈ Fi such that F ⊆ [[ϕ]]M and PRi (ω)(F ) = 1, where [[ϕ]]M = {ω : (M, ω) |= ϕ}. • (M, ω) |= hBi iϕ if there exists a set F ∈ Fi such that F ⊆ [[ϕ]]M and PRi (ω)(F ) > 0. Given a language (set of formulas) L, M is L-measurable if M is appropriate (for some game Γ) and [[ϕ]]M ∈ F for all formulas ϕ ∈ L. It is easy to check that in an L1 -measurable structure, hBi iϕ is equivalent to ¬Bi ¬ϕ. 2 To put our results on iterated admissibility into context, we first consider rationalizability. Pearce [1984] gives two definitions of rationalizability, which give rise to different epistemic characterizations. We repeat the definitions here, using the notation of Osborne and Rubinstein [1994]. Definition 2.1: A strategy σ for player i in game Γ is rationalizable if, for each player j, there is a set Zj ⊆ Σj (Γ) and, for each strategy σ 0 ∈ Zj , a probability measure µσ0 on Σ−j (Γ) whose support is a subset of Z−j such that • σ ∈ Zi ; and • for each player j and strategy σ 0 ∈ Zj , strategy σ 0 is a best response to (the beliefs) µσ0 . The second definition characterizes rationalizability in terms of iterated deletion. Definition 2.2: A strategy σ for player i in game Γ is rationalizable0 if, for each player j, there exists a sequence Xj0 , Xj1 , Xj2 , . . . of sets of strategies for player j such that Xj0 = Σj and, for each strategy ~ k−1 such that σ 0 ∈ Xjk , k ≥ 1, a probability measure µσ0 ,k whose support is a subset of X −j • σ ∈ ∩∞ j=0 Xi ; and • for each player j, each strategy σ 0 ∈ Xjk is a best response to the beliefs µσ0 ,k . Intuitively, Xj1 consists of strategies that are best responses to some belief of player j, and Xjh+1 conh ; that is, sists of strategies in Xjh that are best responses to some belief of player j with support X−j beliefs that assume that everyone else is best reponding to some beliefs assuming that everyone else is responding to some beliefs assuming . . . (h times). Proposition 2.3: [Pearce 1984] A strategy is rationalizable iff it is rationalizable0 . We now give our epistemic characterizations of rationalizability. Let RAT be an abbreviation for RAT 1 ∧. . .∧RAT n ; let Eϕ be an abbreviation of B1 ϕ∧. . .∧Bn ϕ; and define E k ϕ for all k inductively by taking E 0 ϕ to be ϕ and E k+1 ϕ to be E(E k ϕ). Common knowledge of ϕ holds iff E k ϕ holds for all k ≥ 0. We now give an epistemic characterization of rationalizability. Part of the characterization (the equivalence of (a) and (b) below) is well known [Tan and Werlang 1988]; it just says that a strategy is rationalizable iff it can be played in a state where rationality is common knowledge. Theorem 2.4: The following are equivalent: (a) σ is a rationalizable strategy for i in a game Γ; (b) there exists a measurable structure M that is appropriate for Γ and a state ω such that si (ω) = σ and (M, ω) |= E k RAT for all k ≥ 0; 3 (c) there exists a measurable structure M that is appropriate for Γ and a state ω such that si (ω) = σ and (M, ω) |= hBi iE k RAT for all k ≥ 0; (d) there exists a structure M that is appropriate for Γ and a state ω such that si (ω) = σ and (M, ω) |= hBi iE k RAT for all k ≥ 0. Proof: Suppose that σ is rationalizable. Choose Zj ⊆ Σj (Γ) and measures µσ0 for each strategy σ 0 ∈ Zj guaranteed to exist by Definition 2.1. Define an appropriate structure M = (Ω, s, F, PR1 , . . . , PRn ), where • Ω = Z1 × · · · × Zn ; • si (~σ ) = σi ; • F consist of all subsets of Ω; 0 ) otherwise. • PRi (~σ )(~σ 0 ) is 0 if σi0 6= σi and is µσi (σ−i Since each player is best responding to his beliefs at every state, it is easy to see that (M, ~σ ) |= RAT for all states ~σ . It easily follows (formally, by induction on k), that (M, ~σ ) |= E k RAT . Clearly M is measurable. This shows that (a) implies (b). The fact that (b) implies (c) is immediate, since if E k+1 ϕ logically implies Bi E k ϕ, which in turn logically implies hBi ii E k ϕ for all k and all formulas ϕ. The fact that (c) implies (d) is also immediate. Finally, to see that (d) implies (a), suppose that M is a structure appropriate for Γ and ω is a state in M such that si (ω) = σ and (M, ω) |= hBi iE k RAT for all k ≥ 0. For each player j, define the formulas C k inductively by taking Cj0 to be true and Cjk+1 to be RAT j ∧Bj (∧j 0 6=j Cjk0 ). An easy induction shows that for k > 1, Cjk is equivalent to RAT j ∧ Bj (E 0 RAT ∧ . . . ∧ E k−2 RAT ) in appropriate structures. Define Xjk = {sj (ω 0 ) : (M, ω 0 ) |= Cjk }. If σ 0 ∈ Xjk for k ≥ 1, choose some state ω 0 such that (M, ω 0 ) |= RAT j ∧ Bj E k−2 RAT and sj (ω 0 ) = σ 0 , and define µσ0 ,k to be the projection of PRj (ω 0 ) k−1 onto Σ−j . It easily follows that the support of µσ0 ,k is X−j and that σ 0 is a best response with respect to k µσ,k . Finally, since (M, ω) |= hBi iE k RAT for all k ≥ 0, it easily follows that σ = si (ω) ∈ ∩∞ k=0 Xi . 0 Thus, by Definition 2.2, σ is rationalizable and, by Proposition 2.3, σ is rationalizable. We now characterize iterated deletion of strongly dominated (resp., weakly dominated) strategies. Definition 2.5: Strategy σ for player is i strongly dominated by σ 0 with respect to Σ0−i ⊆ Σ−i if ui (σ, τ−i ) > ui (σ, τ−i ) for all τ−i ∈ Σ0−i . Strategy σ for player is i weakly dominated by σ 0 with 0 ) > u (σ, τ 0 ) for some respect to Σ0−i ⊆ Σ−i if ui (σ, τ−i ) ≥ ui (σ, τ−i ) for all τ−i ∈ Σ0−i and ui (σ, τ−i i −i 0 ∈ Σ0 . τ−i −i Strategy σ for player i survives k rounds of iterated deletion of strongly dominated (resp., weakly dominated) strategies if, for each player j, there exists a sequence Xj0 , Xj1 , Xj2 , . . . , Xjk of sets of strategies for player j such that Xj0 = Σj and, if h < k, then Xjh+1 consists of the strategies in Xjh not h , and σ ∈ X k . Strategy σ surstrongly (resp., weakly) dominated by any strategy with respect to X−j i vives iterated deletion of strongly dominated (resp., weakly dominated) strategies if it survives k rounds of iterated deletion for all k. 4 The following well-known result connects strong and weak dominance to best responses. Proposition 2.6: [Pearce 1984] • A strategy σ for player i is not strongly dominated by any strategy with respect to Σ0−i iff there is a belief µσ of player i whose support is a subset of Σ0−i such that σ is a best response with respect to µσ . • A strategy σ for player i is not weakly dominated by any strategy with respect to Σ0−i iff there is a belief µσ of player i whose support is all of Σ0−i such that σ is a best response with respect to µσ . It immediately follows from Propositions 2.3 and 2.6 (and is well known) that a strategy is rationalizable iff it survives iterated deletion of strongly dominated strategies. Thus, the characterization of rationalizability in Theorem 2.4 is also a characterization of strategies that survive iterated deletion of strongly dominated strategies. To characterize iterated deletion of weakly dominated strategies, we need to enrich the langauge L1 somewhat. Let L2 (Γ) be the extension of L1 that includes a primitive proposition play i (σ) for each player i and strategy σ ∈ Σi , and is also closed off under the modal operator ♦. We omit the parenthetical Γ when it is clear from context. We extend the truth relation to L2 in probability structures appropriate for Γ as follows: • (M, ω) |= play i (σ) iff ω ∈ [[σ]]M . • (M, ω) |= ♦ϕ iff there is some structure M 0 appropriate for Γ and state ω 0 such that (M 0 , ω 0 ) |= ϕ. Intuitively, ♦ϕ is true if there is some state and structure where ϕ is true; that is, if ϕi is satisfiable. Note that if ♦ϕ is true at some state, then it is true at all states in all structures. Let play(~σ ) be an abbreviation for ∧nj=1 play j (σj ), and let play −i (σ−i ) be an abbrevation for ∧j6=i play j (σj ). Intuitively, (M, ω) |= play(~σ ) iff s(ω) = σ, and (M, ω) |= play −i (σ−i ) if, at ω, the players other than i are playing strategy profile σ−i . Define the formulas Djk inductively by taking Dj0 to be the formula true, and Djk+1 to be an abbreviation of RAT j ∧ Bj (∧j 0 6=j Djk0 ) ∧ (∧σ−j ∈Σ−j ♦(play −j (σ−j ) ∧ (∧j 0 6=j Djk0 )) ⇒ hBj i(play −j (σ−j )). It is easy to see that Djk implies the formula Cjk defined in the proof of Theorem 2.4, and hence implies RAT j ∧ Bj (E 0 RAT ∧ . . . ∧ E k−2 RAT ). But Djk requires more; it requires that player j k−1 assign positive probability to each strategy profile for the other players that is compatible with D−j . Theorem 2.7: The following are equivalent: (a) the strategy σ for player i survives k rounds of iterated deletion of weakly dominated strategies; 0 0 0 (b) for all k 0 ≤ k, there is a measurable structure M k appropriate for Γ and a state ω k in M k such 0 0 0 0 that si (ω k ) = σ and (M k , ω k ) |= Dik ; 0 0 0 (c) for all k 0 ≤ k, there is a structure M k appropriate for Γ and a state ω k in M k such that 0 0 0 0 si (ω k ) = σ and (M k , ω k ) |= Dik . 5 k In addition, there is a finite structure M = (Ωk , s, F, PR1 , . . . , PRn ) such that Ωk = {(k 0 , i, ~σ ) : 0 0 k 0 k 0 ≤ k, 1 ≤ i ≤ n, ~σ ∈ X1k ×· · ·×Xnk }, s(k 0 , i, ~σ ) = ~σ , F = 2Ω , where Xjk consists of all strategies for player j that survive k 0 rounds of iterated deletion of weakly dominated strategies and, for all states 0 k (k 0 , i, ~σ ) ∈ Ωk , (M , (k 0 , i, ~σ )) |= ∧j6=i Djk . Proof: We proceed by induction on k, proving both the equivalence of (a), (b), and (c) and the existence k of a structure M with the required properties. The result clearly holds if k = 0. Suppose that the result holds for k; we show that it holds for 0 0 0 0 k + 1. We first show that (c) implies (a). Suppose that (M k , ω k ) |= Djk and sj (ω k ) = σj for all k 0 ≤ k+1. It follows that σj is a best response to the belief µσj on the strategies of other players induced by PRk+1 (ω). Since (M k+1 , ω k+1 ) |= Bj (∧j 0 6=j Djk0 ), it follows from the induction hypothesis that j k . Since (M, ω) |= ∧ k the support of µσj is contained in X−j σ−j ∈Σ−j (♦(play −j (σ−j ) ∧ (∧j6=i Dj )) ⇒ k . hBj i(play −j (σ−j ))), it follows from the induction hypothesis that the support of µσj is all of X−j 0 0 0 Since (M k , ω k ) |= Djk for k 0 ≤ k, it follows from the induction hypothesis that σj ∈ Xjk . Thus, σj ∈ Xjk+1 . k+1 = (Ωk+1 , s, F, PR1 , . . . , PRn ). As required, we define We next construct the structure M k+1 0 0 k+1 0 0 Ω = {(k , i, ~σ ) : k ≤ k + 1, 1 ≤ i ≤ n, ~σ ∈ X1k × · · · × Xnk }, s(k 0 , i, ~σ ) = ~σ , F = 2Ω . For a 0 state ω of the form (k 0 , i, ~σ ), since σj ∈ Xjk , by Proposition 2.6, there exists a distribution µk0 ,σj whose k−1 support is all of X−j such that σj is a best response to µσj . Extend µk0 ,σj to a distribution µ0k0 ,i,σj on Ωk+1 as follows: • for i 6= j, let µ0k0 ,i,σj (k 00 , i0 , ~τ ) = µk0 ,σj (~τ−j ) if i0 = j, k 00 = k 0 − 1, and τj = σj , and 0 otherwise; • µ0k0 ,j,σj (k 00 , i0 , ~τ ) = µk0 ,σj (~τ−j ) if i0 = j, k 00 = k 0 , and τj = σj , and 0 otherwise. Let PRj (k 0 , i, ~σ ) = σk0 0 ,i,σj . We leave it to the reader to check that this structure is appropriate. An easy induction on k 0 now shows that (M k+1 0 , (k 0 , i, ~σ )) |= ∧j6=i Djk for i = 1, . . . , n. To see that (a) implies (b), suppose that σj ∈ Xjk+1 . Choose a state ω in M (k + 1, i, ~σ ), where i 6= j. As we just showed, (M Moreover, M k+1 k+1 , (k 0 , i, ~σ ) |= 0 Djk , k+1 and sj of the form (k 0 , i, ~σ ) = σj . is measurable (since F consists of all subsets of Ωk+1 ). Clearly (b) implies (c). Corollary 2.8: The following are equivalent: (a) the strategy σ for player i survives iterated deletion of weakly dominated strategies; (b) there is a measurable structure M that is appropriate for Γ and a state ω such that si (ω) = σ and (M, ω) |= hBi iDik for all k ≥ 0; (c) there is a structure M that is appropriate for Γ and a state ω such that si (ω) = σ and (M, ω) |= hBi iDik for all k ≥ 0. 6 Note that there is no analogue of Theorem 2.4(b) here. This is because there is no state where Dik holds for all k ≥ 0; it cannot be the case that i places positive probability on all strategies (as required by D1k ) and that i places positive probability only on strategies that survive one round of iterated deletion (as required by D2k ), unless all strategies survive one round on iterated deletion. We can say something slightly weaker though. There is some k such that the process of iterated deletion converges; that is, 0 Xjk = Xjk+1 for all j (and hence Xjk = Xjk for all k 0 ≥ k). That means that there is a state where 0 Dik holds for all k 0 > k. Thus, we can show that a strategy σ for player i survives iterated deletion of 0 weakly dominated strategies iff there exists a k and a state ω such that si (ω) = σ and (M, ω) |= Dik for all k 0 > k. Since Cik+1 implies Cik , an anlagous results holds for iterated deletion of strongly dominated 0 0 strategies, with Dik replaced by Cik . It is also worth noting that in a state where Dk holds, an agent does not consider all strategies possible, but only the ones consistent with the appropriate level of rationality. We could require the agent to consider all strategies possible by using LPS’s or nonstandard probability. The only change that this would make to our characterization is that, if we are using nonstandard probability, we would interpret Bi ϕ to mean that ϕ holds with probability infinitesimally close to 1, while hBi iϕ would mean that ϕ holds with probability whose standard part is positive (i.e., non-infinitesimal probability). We do not pursue this point further. 3 Strong Admissibility We have formalized iterated admissibility by saying that an agent consider possible all strategies consistent with the appropriate rationality assumption. But why focus just on strategies? We now consider a stronger admissibility requirement that we call, not surprisingly, strong admissibility. Here we require, intuitively, that all an agent knows about the other agents is that they satisfy the appropriate rationality assumptions. Thus, the agent ascribes positive probability to all beliefs that the other agents could have as well as all the strategies they could be using. By considering strong admissibility, we will be able to relate work on “all I know” [Halpern and Lakemeyer 2001; Levesque 1990], BFK’s notion of complete structures, and admisibility. Roughly speaking, we interpret “all agent i knows is ϕ” as meaning that agent i believes ϕ, and considers possible every formula about the other players’ strategies and beliefs consistent with ϕ. Thus, what “all I know” means is very sensitive to the choice of language. Let L0 be the language whose only formulas are (Boolean combinations of) formulas of the form play i (σ), i = 1, . . . , n, σ ∈ Σi . Let L0i consist of just the formulas of the form play i (σ), and let L0−i = ∪j6=i L0j . Define Oi− ϕ to be an abbreviation for Bi ϕ ∧ (∧ψ∈L0 ♦(ϕ ∧ ψ) ⇒ hBi iψ). Then it is easy to see that Djk+1 is just RAT j ∧ Oj− (∧j 0 6=j Djk0 ). −i We can think of Oi− ϕ as saying “all agent i knows with respect to the language L0 is ϕ.” The language L0 is quite weak. To relate our results to those of BFK, even the language L2 is too weak, since it does not allow an agent to express probabilistic beliefs. Let L3 (Γ) be the language that extends L2 (Γ) by allowing formulas of the form pr i (ϕ) ≥ α and pr i (ϕ) > α, where α is a rational number in [0, 1]; pr i (ϕ) ≥ α can be read as “the probability of ϕ according to i is at least α”, and similarly for pr i (ϕ) > α. We allow nesting here, so that we can have a formula of the form pr j (play i (σ) ∧ pr k (play i (σ 0 )) > 1/3) ≥ 1/4. As we would expect, • (M, ω) |= pr i (ϕ) iff PRi (ω)([[ϕ]]M ) ≥ α. 7 The restriction to α being rational allows the language to be countable. However, as we now show, it is not too serious a restriction. Let L4 (Γ) be the language that extends L2 (Γ) by closing off under countable conjunctions, so that if ϕ1 , ϕ2 , . . . are formulas, then so is ∧∞ m=1 ϕm , and formulas of the form pr i (ϕ) > α, where α is a real number in [0, 1]. (We can express pr i (ϕ) ≥ α as the countable conjunction ∧β<α,β∈Q∩[0,1] pr i (ϕ) > β, where Q is the set of rational numbers, so there is no need to include formulas of the form pr i (ϕ) ≥ α explicitly in L4 (Γ).) We omit the parenthetical Γ in L3 (Γ) and L4 (Γ) when the game Γ is clear from context. A subset Φ of L3 is L3 -realizable if there exists an appropriate structure M for Γ and state ω in M such that, for all formulas ϕ ∈ L3 , (M, ω) |= ϕ iff ϕ ∈ Φ.1 We can similarly define what it means for a subset of L4 to be L4 -realizable. Lemma 3.1: Every L3 -realizable set can be uniquely extended to an L4 -realizable set. Proof: It is easy to see that every L3 -realizable set can be extended to an L4 -realizable set. For suppose that Φ is L3 -realizable. Then there is some state ω and structure M such that, for every formula ϕ ∈ L3 , we have that (M, ω) |= ϕ iff ϕ ∈ Φ. Let Φ0 consist of the L4 formulas true at ω. Then clearly Φ0 is an L4 -realizable set that extends Φ. To show that the extension is unique, suppose that there are two L4 -realizable sets, say Φ1 and Φ2 , that extend Φ. We want to show that Φ1 = Φ2 . To do this, we consider two language, L5 and L6 , intermediate between L3 and L4 . Let L5 be the language that extends L2 by closing off under countable conjunctions and formulas of the form pr i (ϕ) > α, where α is a rational number in [0, 1]. Thus, in L5 , we have countable conjunctions and disjunctions, but can talk explicitly only about rational probabilities. Nevertheless, it is easy to see that for every formula ϕ ∈ L4 , there is an formula equivalent formula ϕ0 ∈ L5 , since if α is a real number, then pr i (ϕ) > α is equivalent to ∨β>α, β∈[0,1]∩Q pr i (ϕ) > β (an infinite disjunction ∞ ∨∞ i=1 ϕi can be viewed as an abbreviation for ¬ ∧i=1 ¬ϕi ). Next, let L6 be the result of closing off formulas in L3 under countable conjunction and disjunction. Thus, in L6 , we can apply countable conjunction and disjunction only at the outermost level, not inside the scope of pr i . We claim that for every formula ϕ ∈ L5 , there is an equivalent formula in L6 . More precisely, for every formula ϕ ∈ L5 , there exist formulas ϕij ∈ L3 , 1 ≤ i, j < ∞ such that ϕ is ∞ equivalent to ∧∞ m=1 ∨n=1 ϕmn . We prove this by induction on the structure of ϕ. If ϕ is RAT i , play i (σ), or true, then the statement is clearly true. The result is immediate from the induction hypothesis if ϕ is a countable conjunction. If ϕ has the form ¬ϕ0 , we apply the induction hypothesis, and observe ∞ ∞ ∞ that ¬(∧∞ m=1 ∨n=1 ϕmn ) is equivalent to ∨m=1 ∧n=1 ¬ϕmn . We can convert this to a conjunction of disjunctions by distributing the disjunctions over the conjunctions in the standard way (just as (E1 ∩ E2 ) ∪ (E3 ∩ E4 ) is equivalent to (E1 ∪ E3 ) ∩ (E1 ∪ E4 ) ∩ (E2 ∪ E3 ) ∩ (E2 ∪ E4 )). Finally, if ϕ has ∞ the form pr i (ϕ0 ) > α, we apply the induction hypothesis, and observe that pr i (∧∞ m=1 ∨n=1 ϕmn ) > α is equivalent to ∞ M N 0 ∨α0 >α,α0 ∈Q∩[0,1] ∧∞ M =1 ∨N =1 pr i (∧m=1 ∨n=1 ϕmn ) > α . The desired result follows, since if two states agree on all formulas in L3 , they must agree on all formulas in L6 , and hence on all formulas in L5 and L4 . 1 For readers familiar with standard completeness proofs in modal logic, if we had axiomatized the logic we are implicitly using here, the L3 -realizable sets would just be the maximal consistent sets in the logic. 8 The choice of language turns out to be significant for a number of our results; we return to this issue at various points below. With this background, we can define strong admissibility. Let L3i consist of all formulas in L3 of the form pr i (ϕ) ≥ α and pr i (ϕ) > α (ϕ can mention pr i ; it is only the outermost modal operator that must be i). Intuitively, L3i consists of the formulas describing i’s beliefs. Let L3i+ consist of L3i together with formulas of the form true, RAT i , and play i (σ), for σ ∈ Σi . Let L3(−i)+ be an abbreviation for ∪j6=i L3j+ . We can similarly define L4i and L4i+ . If ϕ ∈ L3(−i)+ , define Oi ϕ, read “all agent i knows (with respect to L3 ) is ϕ,” as an abbreviation for the L4 formula Bi ϕ ∧ (∧ψ∈L3 ♦(ϕ ∧ ψ) ⇒ hBj iψ). (−i)+ Thus, Oi ϕ holds if agent i believes ϕ but does not know anything beyond that; he ascribes positive probability to all formulas in L3(−i)+ consistent with ϕ. This is very much in the spirit of the HalpernLakemeyer [2001] definition of Oi in the context of epistemic logic. Of course, we could go further and define a notion of “all i knows” for the language L4 . Doing this would give a definition that is even closer to that of Halpern and Lakemeyer. Unfortunately, we cannot require than agent i ascribe positive probability to all the formulas in L4(−i)+ consistent with ϕ; in general, there will be an uncountable number of distinct and mutually exclusive formulas consistent with ϕ, so they cannot all be assigned positive probability. This problem does not arise with L3 , since it is a countable language. Halpern and Lakemeyer could allow an agent to consider an uncountable set of worlds possible, since they were not dealing with probabilistic systems. This stresses the point that the notion of “all I know” is quite sensitive to the choice of language. Define the formulas Fik inductively by taking Fi0 to be the formula true, and Fik+1 to an abbreviation of RAT i ∧ Oi (∧j6=i Fjk ). Thus, Fjk+1 says that i is rational, believes that all the other players satisfy level-k rationality (i.e., Fjk ), and that is all that i knows. An easy induction shows that Fjk+1 implies that j is rational and j believes that everyone believes (k times) that everyone is rational. Moreover, it is easy to see that Fjk+1 implies Djk+1 . The difference is that instead of requiring just that j assign positive k , it requires that j assign positive probability to probability to all strategy profiles compatible with F−j k all formulas compatible with F−j . A strategy σi for player i is kth-level strongly admissible if it is consistent with Fik ; that is, if play i (σi ) ∧ Fik is satisfied in some state. The next result shows that strong admissibility characterizes iterated deletion, just as admissibility does. Theorem 3.2: The following are equivalent: (a) the strategy σ for player i survives k rounds of iterated deletion of weakly dominated strategies; 0 0 0 (b) for all k 0 ≤ k, there is a measurable structure M k appropriate for Γ and a state ω k in M k such 0 0 0 0 that si (ω k ) = σ and (M k , ω k ) |= Fik ; 0 0 0 (c) for all k 0 ≤ k, there is a structure M k appropriate for Γ and a state ω k in M k such that 0 0 0 0 si (ω k ) = σ and (M k , ω k ) |= Fik ; Proof: The proof is similar in spirit to the proof of Theorem 2.7. We again proceed by induction on k. The result clearly holds for k = 0. If k = 1, the proof that (c) implies (a) is essentially identical to that of Theorem 2.7; we do not repeat it here. 9 To prove that (a) implies (b), we need the following three lemmas; the first shows that a formula is always satisfied in a state that has probability 0; the second shows that that we can get a new structure with a world where agent i ascribes positive probability to each of a countable collection of satisfiable formulas in L3−i ; and the third shows that formulas in L4i+ for different players i are independent; that is, if ϕi ∈ L4i+ is satisfiable, then so is ϕ1 ∧ . . . ∧ ϕn . Lemma 3.3: If ϕ ∈ L4 is satisfiable in a measurable structure, then there exists a measurable structure M and state ω such that (M, ω) |= ϕ, {ω} is measurable, PRj (ω)({ω}) = 0 for j = 1, . . . , n. Proof: Suppose that (M 0 , ω 0 ) |= ϕ, where M 0 = (Ω0 , s0 , F 0 , PR01 , . . . , PR0n ). Let Ω = Ω0 ∪ {ω}, where where ω is a fresh state; let F be the smallest σ-algebra that contains F and {ω}; let s and PRj agree with s0 and PR0j when restricted to states in Ω0 (more precisely, if ω 00 ∈ Ω0 , then PRj (ω 00 )(A) = PR0j (ω 00 )(A ∩ Ω0 ) for j = 1, . . . , n). Finally, define si (ω) = si (ω 0 ), and take PRj (ω)(A) = PR0j (ω 0 )(A ∩ Ω0 ) for j = 1, . . . , n. Clearly {ω} is measurable, and PRj (ω)({ω}) = 0 for j = 1, . . . , n. An easy induction on structure shows that for all formulas ψ, (a) (M, ω) |= ψ iff (M, ω 0 ) |= ψ, and (b) for all states ω 00 ∈ Ω0 , we have that (M, ω 00 ) |= ψ iff (M 0 , ω 00 ) |= ψ. It follows that (M, ω) |= ϕ, and that M is measurable. ~ Φ0 is a countable collection of formulas in L4 , ϕ ∈ L4 , and Σ0 Lemma 3.4: Suppose that ~σ ∈ Σ, −i −i −i is a set of strategy profiles in Σ−i such that (a) for each formula ϕ0 ∈ Φ0 , there exists some profile σ−i ∈ Σ0−i such that ϕ ∧ ϕ0 ∧ play −i (σ−i ) is satisfied in a measurable structure, and (b) for each profile σ−i ∈ Σ0−i , play −i (σ−i ) is one of the formulas in Φ0 . Then there exists a measurable structure M and state ω such that s(ω) = ~σ , (M, ω) |= play −j (σ−i ) ≥ α iff µj (σ−i ) ≥ α (that is, µ−i agrees with PRi (ω) when marginalized to strategy profiles in Σ0−i ), and (M, ω) |= Bi ϕ ∧ hBi iϕ0 for all ϕ0 ∈ Φ0 . Proof: Let Φ0 and Σ0−i be as in the statement of the lemma. Suppose that Φ0 = {ϕ1 , ϕ2 , . . . , . . .}. By 0 ∈ Σ0 , measurable strucassumption, for each formula ϕk ∈ Φ0 , there exists some strategy profile σ−i −i k k 0 ), k k k k k k k k ture M = (Ω , s , F , PR1 , . . . , PRn ), and ω ∈ Ω such that (M , ω ) |= ϕ ∧ ϕk ∧ play −i (σ−i for k = 1, 2, . . .. By Lemma 3.3, we can assume without loss of generality that {ω k } ∈ F k and ∞ PRkj (ω k )({ω i }) = 0. Define M ∞ = (Ω∞ , s∞ , F ∞ , PR∞ 1 , . . . , PRn ) as follows: k • Ω∞ = ∪∞ k=0 Ω ∪ {ω}, where ω is a fresh state; • F ∞ is the smallest σ-algebra that contains {ω} ∪ F1 ∪ F2 ∪ . . .; k ∞ • s∞ agrees with PRkj when restricted to states in Ωk , except that s∞ σ; i (ω ) = σi and s (ω) = ~ k k 0 k • PR∞ j agrees with PRj when restricted to states in Ω (more precisely, if ω ∈ Ω , then k ∞ ∞ 1 0 0 k 2 PR∞ j (ω )(A) = PRj (ω )(A∩Ω ), except that PRi (ω) = PRi (ω ) = PRi (ω ) = · · · is de1 2 fined to be a distribution with support {ω , ω , . . .} (so that all these states are given positive prob∞ ability) such that PR∞ i (ω) agrees µ when marginalized to profiles in Σ−i , and PRj (ω)({ω}) = 1 for j 6= i. It is easy to see that our assumptions guarantee that this can be done. We can now prove by a straightforward induction on the structure of ψ that (a) for all formulas ψ, k = 1, 2, 3, . . ., and states ω 0 ∈ Ωk − {ω k }, we have that (M k , ω 0 ) |= ψ iff (M ∞ , ω 0 ) |= ψ; and 10 (b) for all formulas ψ ∈ L4(−i)+ , k = 1, 2, 3, . . ., and (M k , ω k ) |= ψ iff (M ∞ , ω k ) |= ψ. (Here it is k k important that PR∞ j (ω ) = PRj (ω) = 0 for j 6= i; this ensures that j’s beliefs about i’s strategies and k ∞ k k k beliefs unaffected by the fact that ski (ω k ) 6= s∞ i (ω ) and PRi (ω ) 6= PRi (ω ).) It easily follows that (M ∞ , ω) |= Bi ϕ ∧ hBi iϕ0 for all ϕ0 ∈ Φ0 . Lemma 3.5: If ϕi ∈ L4i+ is satisfiable for i = 1, . . . , n, then ϕ1 ∧ . . . ∧ ϕn is satisfiable. Proof: Suppose that (M i , ω i ) |= ϕi , where M i = (Ωi , si , F i , PRi1 , . . . , PRin ) and ϕi ∈ L4i+ . By Lemma 3.3, we again assume without loss of generality that {ω i } ∈ F i and PRj (ω i )({ω i }) = 0. Let M ∗ = (Ω∗ , s∗ , F ∗ , PR∗1 , . . . , PR∗n ), where • Ω∗ = ∪ni=1 Ωi ; • F ∗ is the smallest σ-algebra containing F 1 ∪ . . . ∪ F n ; • s∗ agrees with sj on states in Ωj except that s∗i (ω j ) = sii (ω i ) (so that s∗ (ω 1 ) = · · · = s∗ (ω n )); • PR∗i agrees with PRji on states in Ωj except that PR∗i (ω j ) = PRii (ω i ) (so that PR∗i (ω 1 ) = · · · = PR∗i (ω n ) = PRii (ω i )). We can now prove by induction on the structure of ψ that (a) for all formulas ψ, i = 1, . . . , n, and states ω 0 ∈ Ωi , we have that (M i , ω 0 ) |= ψ iff (M ∗ , ω 0 ) |= ψ; (b) for all formulas ψ ∈ L4i+ , 1 ≤ i, j ≤ n, (M i , ω i ) |= ψ iff (M ∗ , ω j ) |= ψ (again, here it is important that PR∗i (ω j ) = 0 for j = 1, . . . , n). Note that part (b) implies that the states ω 1 , . . . , ω n satisfy the same formulas in M ∗ . It easily follows that (M ∗ , ω i ) |= ϕ1 ∧ . . . ∧ ϕn for i = 1, . . . , n. We can now prove the theorem. Again, let Xjk be the strategies for player j that survive k rounds of iterated deletion of weakly dominated strategies. To see that (a) implies (b), suppose that σi ∈ Xik+1 . k such that σ is a best response By Proposition 2.6, there exists a distribution µi whose support is X−i i k , and all j 6= i, the formula to µi . By the induction hypothesis, for each strategy profile τ−i ∈ X−i play j (τj ) ∧ Fj0 is satisfied in a measurable structure. By Lemma 3.5, play −j (τ−j ) ∧ (∧j6=i Fjk ) is satisfied in a measurable structure. Taking ϕ to be ∧j6=i Fjk , by Lemma 3.4, there exists a measurable k is µ , s (ω) is σ , and (M, ω) |= structure M and state ω in M such that the marginal of PRi (ω) on X−i i i i ♦(ψ ∧ (∧j6=i Fjk )) ⇒ hBj iψ). It follows that (M, ω) |= RAT i , and hence Bi (∧j6=i Fjk ) ∧ (∧ψ∈L3 (−j)+ that (M, ω) |= Fik+1 , as desired. It is immediate that (b) implies (c). Corollary 3.6: The following are equivalent: (a) the strategy σ for player i survives iterated deletion of weakly dominated strategies; (b) there is a measurable structure M that is appropriate for Γ and a state ω such that si (ω) = σ and (M, ω) |= hBi iFik for all k ≥ 0; (c) there is a structure M that is appropriate for Γ and a state ω such that si (ω) = σ and (M, ω) |= hBi iFik for all k ≥ 0. Proof: The proof is essentially identical to that of Corollary 2.8, so is omitted here. 11 4 4.1 Complete and Canonical Structures Canonical Structures Intuitively, to check whether a formula is strongly admissible, and, more generally, to check if all agent i knows is ϕ, we want to start with a very rich structure M that contains all possible consistent sets of formulas, so that if ϕ ∧ ψ is satisfied at all, it is satisfied in that structure. Motivated by this intuition, Halpern and Lakemeyer [2001] worked in the canonical structure for their language, which contains a state corresponding to every consistet set of formulas. We do the same thing here. Define the canonical structure M c = (Ωc , sc , F c , PRc1 , . . . , PRcn ) for L4 as follows: • Ωc = {ωΦ : Φ is a realizable subset of L4 (Γ}; • sc (ωΦ ) = ~σ iff play(σ) ∈ Φ; • F c = {Fϕ : ϕ ∈ L4 }, where Fϕ = {ωΦ : ϕ ∈ Φ}; • Prci (ωΦ )(Fϕ ) = inf{α : pr i (ϕ) > α ∈ Φ}. Lemma 4.1: M c is an appropriate measurable structure for Γ. Proof: It is easy to see that F c is a σ-algebra, since the complement of Fϕ is F¬ϕ and ∩∞ m=1 Fϕi = F∧ ∞ . Given a strategy σ for player i, [[σ]]M c = Fplay i (σ ) ∈ F. Moreover, each realizable m=1 ϕm set Φ that includes play i (σ) must also include pr i (play i (σ)) = 1, so that PRi (ωΦ )(si (ωΦ )) = PRi (ωΦ )(Fplay i (si (ωΦ ) ) = 1. Similarly, suppose that PRi (ωΦ ) = π. Then {ω ∈ Ωc : PRi (ω) = π} = ∩ϕ∈L3 ∩{α∈Q∩[0,1]:π([[ϕ]]M c )≥α} Fϕ≥α ∈ F c . Moreover, if α ∈ Q ∩ [0, 1], then π([[ϕ]]M c ) ≥ α iff pr i (ϕ) ≥ α ∈ Φ. But if pr i (ϕ) ≥ α ∈ Φ, then pr i (pr i (ϕ) ≥ α) = 1 ∈ Φ. It easily follows that PRi (ωΦ )({ω : PRi (ω) = π}) = 1. Finally, the definition of F c guarantees that every set [[ϕ]]M c is measurable and that PRi (ωΦ ) is indeed a probability distribution on (Ωc , F c ). The following result is the analogue of the standard “truth lemma” in completeness proofs in modal logic. Proposition 4.2: For ψ ∈ L4 , (M c , ωΦ ) |= ψ iff ψ ∈ Φ. Proof: A straightforward induction on the structure of ψ. We have constructed a canonical structure for L4 . It follows easily from Lemma 3.1 that the canonical structure for L3 (where the states are realizable L3 sets) is isomorphic to M c . (In this case, the set F c of measurable sets would be the smallest σ-algebra containing [[ϕ]]M for ϕ ∈ L3 .) Thus, the choice of L3 vs. L4 does not play an important role when constructing a canonical structure. A strategy σi for player i survives iterated deletion of weakly dominated strategies iff the L4 formula k undominated (σi ) = play i (σi ) ∧ (∧∞ k=1 hBi iFi ) is satisfied at some state in the canonical structure. But there are other structures in which undominated (σi ) is satisfied. One way to get such a struture is by essentially “duplicating” states in the canonical structure. The canonical structure can be embedded in a structure M if, for all L3 -realizable sets Φ, there is a state ωΦ in M such that (M, ωΦ ) |= ϕ iff 12 ϕ ∈ Φ. Clearly undominated (σi ) is satisfied in any structure in which the canonical structure can be embedded. A structure in which the canonical structure can be embedded is in a sense larger than the canonical structure. But undominated (σi ) can be satisfied in structures smaller than the canonical structure. (Indeed, with some effort, we can show that it is satisfiable in a structure with countably many states.) There are two reasons for this. The first is that to satisfy undominated (σi ), there is no need to consider a structure with states where all the players are irrational. It suffices to restrict to states where at least one player is using a strategy that survives at least one round of iterated deletion. This is because players know their strategy; thus, in a state where a strategy σj for player j is admissible, player j must ascribe positive probability to all other strategies; however, in those states, player j still plays σj . A perhaps more interesting reason that we do not need the canonical structure is our use of the language L3 . Strong admissibility guarantees that player j will ascribe positive probability to all formulas ϕ consistent with rationality. Since a finite conjunction of formulas in L3 is also a formula in L3 , player j will ascribe positive probability to all finite conjunctions of formulas consistent with rationality. But a state is characterized by a countable conjunction of formulas. Since L3 is not closed under countable conjunctions, a structure that satisfies undominated (σi ) may not have states corresponding to all L3 -realizable sets of formulas. If we had used L4 instead of L3 in the definition of strong admissibility (ignoring the issues raised earlier with using L4 ), then there would be a state corresponding to every L4 -realizable (equivalently, L3 -realizable) set of formulas. Alternatively, if we consider appropriate structures that are compact in a topology where all sets definable by formulas (i.e., sets of the form [[ϕ]]M , for ϕ ∈ L3 ) are closed (in which case they are also open, since [[¬ϕ]]M is the complement of [[ϕ]]M ), then all states where at least one player is using a strategy that survives at least one round of iterated deletion will be in the structure. Although, as this discussion makes clear, the formula that characterizes strong admissibility can be satisfied in structures quite different from the canonical structure, the canonical structure does seem to be the most appropriate setting for reasoning about statements involving “all agent i knows”, which is at the heart of strong admissibility. Moreover, as we now show, canonical structures allow us to relate our approach to that of BFK. 4.2 Complete Structures BFK worked with complete structures. We now want to show that M c is complete, in the sense of BFK. To make this precise, we need to recall some notions from BFK (with some minor changes to be more consistent with our notation). BFK considered what they called interactive probability structures. These can be viewed as a special case of probability structures. A BFK-like structure (for a game Γ) is a probability structure M = (Ω, s, F, PR1 , . . . , PRn ) such that there exist spaces T1 , . . . , Tn (where Ti can be thought of as the type space for player i) such that ~ × T~ , via some isomorphism h; • Ω is isomorphic to Σ • if h(ω) = ~σ × ~t, then – s(ω) = ~σ , 13 – taking Ti (ω) = ti (i.e., the type of player i in h(ω) is ti ); the support of PRi (ω) is contained in {ω 0 : si (ω 0 ) = σ 0 , Ti (ω 0 ) = ti }, so that PRi (ω) induces a probability on Σ−i × T−i ; – PRi (ω) depends only on Ti (ω), in the sense that if Ti (ω) = Ti (ω 0 ), then PRi (ω) and PRi (ω 0 ) induce the same probability distribution on Σ−i × T−i . ~ × T~ is complete if, for every for each A BFK-like structure M whose state space is isomorphic to Σ distribution µi over Σ−i ×T−i , there is a state ω in M such that the probability distribution on Σ−i ×T−i induced by PRi (ω) is µi . Proposition 4.3: M c is complete BFK-like structure. Proof: A set Φ ⊆ L3i is L3i -realizable if there exists an appropriate structure M for Γ and state ω in M such that, for all formulas ϕ ∈ L3 , (M, ω) |= ϕ iff ϕ ∈ Φ. Take the type space Ti to consist of all ~ × T~ , where Ti (ω) is the L3i -realizable sets of formulas. There is an isomorphism h between Ωc and Σ i-realizable type of formulas of the form pr i (ϕ) ≥ α that are true at ω; that is, h(ω) = s(ω) × T1 (ω) × · · · × Tn (ω). It follows easily from Lemma 3.5 that h is a surjection. we can identify Ωc , the state space ~ × T~ . in the canonical structure, with Σ To prove that M c is complete, given a probability µ on Σ−i × T−i , we must show that there is some state ω in M c such that the probability induced by PRi (ω) on Σ−i × T−i is µ. Let M µ = (Ωsigma,µ , F µ , sµ , PRµ1 , . . . , PRµn ), where M µ are defined as follows: • Ωµ = Ωc ∪ Σ × {µ} × T−i ; • F µ is the smallest σ-algebra that contains F c and all sets of the form ~σ × {µ} × [[ϕ]]0M c , and [[ϕ]]0M c consists of the all type profiles t−i such that, for some state ω in M c , (M c , ϕ) |= ϕ and T−i (ϕ) = t−i ; • sµ (ω) = sc (ω) for ω ∈ Ωc , and sµ (~σ × {µ} × ~t) = ~σ ; • PRµj (ω) = PRcj (ω) for ω ∈ Ωc , j = 1, . . . , n; for j 6= i, PRµj (~σ × µ × t−i ) = PRj (ω), where sj (ω) = σj and Tj (ω) = tj (this is well defined, since if sj (ω 0 ) = σj and Tj (ω 0 ) = tj , then PRj (ω) = PRj (ω 0 ); finally, PRµi (~σ × µ × t−i ) is a distribution whose support is contained in {σi } × Σ−i × {µ} × T−i , and PRµi (~σ × µ × t−i )(~σ × µ × [[ϕ]]0M c ) = µ([[ϕ]]0M c ). ~ × {µ} × T−i . The construction of M µ guarantees that for ϕ ∈ L4 Choose an arbitrary state ω ∈ Σ (−i)+ , µ 0 c 0 (M , ω) |= pr i (ϕ) > α iff µ([[ϕ]]M c ) > α. By the construction of M , there exists a state ω ∈ Ωc such that (M c , ω 0 ) |= ψ iff (M µ , ω) |= ψ. Thus, the distribution on Σ−i × T−i induced by PRi (ω) is µ, as desired. This shows that M c is complete. We now would like to show that every measurable complete BFK-like structure is the canonical model. This is not quite true because states can be duplicated in an interactive structure. This suggests that we should try to show that the canonical structure can be embedded in every measurable complete structure. We can essentially show this, except that we need to restrict to strongly measurable complete structures, where a structure is strongly measurable if it is measurable and the only measurable sets are those defined by L4 formulas (or, equivalently, the set of measurable sets is the smallest set that contains the sets defined by L3 formulas). We explain where strong measurability is needed at the end of the proof of the following theorem. 14 Theorem 4.4: If M is a strongly measurable complete BFK-like structure, then the canonical structure can be embedded in M . Proof: Suppose that M is a strongly measurable complete BFK-like structure. We can assume without ~ × T~ . To prove the result, we need the loss of generality that the state space of M has the form Σ following lemmas. Lemma 4.5: If M is BFK-like, the truth of a formula ϕ ∈ L4i at a state ω in M depends only on i’s type; That is, if Ti (ω) = Ti (ω 0 ), then (M, ω) |= ϕ iff (M, ω 0 ) |= ϕ. Similarly, the truth of a formula in Li+ in ω depends only on si (ω) and Ti (ω), and the truth of a formula in L4i+ in ω depends only on T−i (ω). Proof: A straightforward induction on structure. Define a basic formula to be one of the form ψ1 ∧ . . . ∧ ψn , where ψi ∈ L3i+ for i = 1, . . . , n. Lemma 4.6: Every formula in L3 is equivalent to a finite disjunction of basic formulas. Proof: A straightforward induction on structure. Lemma 4.7: Every formula in L3i+ is equivalent to a disjunction of formulas of the form play i (σ)∧(¬)RATi ∧(¬)pr i (ϕ1 ) > α1 ∧. . .∧(¬)pr i (ϕm ) > αm ∧(¬)pr i (ψ1 ) ≥ β1 ∧. . .∧(¬)pr i (ψm0 ) ≥ βm0 , (1) where ϕ1 , . . . , ϕm , ψ1 , . . . , ψm0 ∈ L3(−i)+ and the “(¬)” indicates that the presence of negation is optional. Proof: A straightforward induction on the structure of formulas, using the observation that ¬play i (σ) is equivalent to ∨{sigma0 ∈Σi :σ0 6=σ} play i (σ 0 ). Lemma 4.8: If ϕ ∈ L3 is satisfiable, then [[ϕ]]M 6= ∅. Proof: By Lemma 4.6, it suffices to prove the result for the case that ϕ is a basic formula. By Lemma 4.7, it suffices to assume that the the “i-component” of the basic formula is a conjunction. We now prove the result by induction on the depth of nesting of the modal operator pr i in ϕ. (Formally, define D(ψ), the depth of nesting of pr i ’s in ψ, by induction on the structure of ψ. if ψ has the form play j (σ), RAT j , or true, then D(ψ) = 0; D(¬ψ) = D(ψ); D(ψ1 ∧ ψ2 ) = max(D(ψ1 ), D(ψ2 )); and D(pr i (ψ) > α) = D(pr i (ψ) ≥ α) = 1 + D(ψ).) Because the state space Ω of M is essentially a product space, by Lemma 4.5, it suffices to prove the result for formulas in L3(i)+ . It is clear that ϕ possibly puts constraints on what strategy i is using, the probability of strategy profiles in Σ−i , and the probability of formulas that appear in the scope of pr i in ϕ. If M 0 = (Ω0 , s0 , F 0 , PR01 , . . . , PR0n ) is a structure and ω 0 ∈ Ω0 , then (M 0 , ω 0 ) |= ϕ iff s0i (ω 0 ) and PR0i (ω 0 ) satisfies these constraints. (We leave it to the reader to formalize this somewhat informal claim.) By the induction hypothesis, each formula in the scope of pr i in ϕ that is assigned positive probability by PRi (ω 0 ) is satisfied in M . Since M is complete and 15 measurable, there is a state ω in M such that si (ω) = s0i (ω 0 ) and PRi (ω) places the same constraints on formulas that appear in ϕ as PRi . We must have (M, ω) |= ϕ. Returning to the proof of the theorem, suppose that M = (Ω, s, F, PR1 , . . . , PRn ). Given a state ω ∈ Ωc , we claim that there must be a state ω 0 in M such that s(ω 0 ) = sc (ω) and, for all i = 1, . . . , n, PRci (ω)([[ψ]]M c ) = PRi (ω 0 )([[ψ]]M ). to show this, because of Ω is a product space, and PRi (ω 0 ) depends only on Ti (ω 0 ), it suffices to show that, for each i, there exists a state ωi in M such that, for each i, PRci (ω)([[ψ]]M c ) = PRi (ωi )([[ψ]]M ). By Lemma 4.8, if [[ψ]]M c 6= ∅, then [[ψ]]M 6= ∅. Thus, the existence of ωi follows from the assumption that M is complete and strongly measurable. Roughly speaking, To understand the need for strong measurability here, note that even without strong measurability, the argument above tells us that there exists an appropriate measure defined on sets of the form [[ϕ]]M for ϕ in L3(−i)+ . We can easily extend µ to a measure µ0 on sets of the form [[ϕ]]M for ϕ in L4(−i)+ . However, if the set F of measurable sets in M is much richer than the sets definable by L4 formulas, it is not clear that we can extend µ0 to a measure on all of F. In general, a countably additive measure defined on a subalgebra of a set F of measurable sets cannot be extended to F. For example, it is known that, under the continuum hypothesis, Lebesgue measure defined on the Borel sets cannot be extended to all subsets of [0, 1] [Ulam 1930]; see [Keisler and Tarski 1964] for further discussion). Strong measurability allows us to avoid this problem. References Brandenburger, A., A. Friedenberg, and J. Keisler (2008). Admissibility in games. Econometrica 76(2), 307–352. Halpern, J. Y. and G. Lakemeyer (2001). Multi-agent only knowing. Journal of Logic and Computation 11(1), 41–70. Keisler, J. and A. Tarski (1964). From accessible to inaccessible cardinals. Fundamenta Mathematica 53, 225–308. Levesque, H. J. (1990). All I know: a study in autoepistemic logic. Artificial Intelligence 42(3), 263–309. Mas-Colell, A., M. Whinston, and J. Green (1995). Microeconomic Thoery. Oxford, U.K.: Oxford University Press. Osborne, M. J. and A. Rubinstein (1994). A Course in Game Theory. Cambridge, Mass.: MIT Press. Pearce, D. G. (1984). Rationalizable strategic behavior and the problem of perfection. Econometrica 52(4), 1029–1050. Rajan, U. (1998). Trembles in the Bayesian foundation of solution concepts. Journal of Economic Theory 82, 248–266. Tan, T. and S. Werlang (1988). The Bayesian foundation of solution concepts of games. Journal of Economic Theory 45(45), 370–391. Ulam, S. (1930). Zur masstheorie in der allgemeinen mengenlehre. Fundamenta Mathematicae 16, 140–150. 16